xref: /linux/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 0d456bad36d42d16022be045c8a53ddbb59ee478)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/delay.h>
37 #include <linux/moduleparam.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 
41 #include <linux/ip.h>
42 #include <linux/tcp.h>
43 
44 #include "ipoib.h"
45 
46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
47 static int data_debug_level;
48 
49 module_param(data_debug_level, int, 0644);
50 MODULE_PARM_DESC(data_debug_level,
51 		 "Enable data path debug tracing if > 0");
52 #endif
53 
54 static DEFINE_MUTEX(pkey_mutex);
55 
56 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
57 				 struct ib_pd *pd, struct ib_ah_attr *attr)
58 {
59 	struct ipoib_ah *ah;
60 	struct ib_ah *vah;
61 
62 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
63 	if (!ah)
64 		return ERR_PTR(-ENOMEM);
65 
66 	ah->dev       = dev;
67 	ah->last_send = 0;
68 	kref_init(&ah->ref);
69 
70 	vah = ib_create_ah(pd, attr);
71 	if (IS_ERR(vah)) {
72 		kfree(ah);
73 		ah = (struct ipoib_ah *)vah;
74 	} else {
75 		ah->ah = vah;
76 		ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
77 	}
78 
79 	return ah;
80 }
81 
82 void ipoib_free_ah(struct kref *kref)
83 {
84 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
85 	struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
86 
87 	unsigned long flags;
88 
89 	spin_lock_irqsave(&priv->lock, flags);
90 	list_add_tail(&ah->list, &priv->dead_ahs);
91 	spin_unlock_irqrestore(&priv->lock, flags);
92 }
93 
94 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
95 				  u64 mapping[IPOIB_UD_RX_SG])
96 {
97 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
98 		ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
99 				    DMA_FROM_DEVICE);
100 		ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
101 				  DMA_FROM_DEVICE);
102 	} else
103 		ib_dma_unmap_single(priv->ca, mapping[0],
104 				    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
105 				    DMA_FROM_DEVICE);
106 }
107 
108 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
109 				   struct sk_buff *skb,
110 				   unsigned int length)
111 {
112 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
113 		skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
114 		unsigned int size;
115 		/*
116 		 * There is only two buffers needed for max_payload = 4K,
117 		 * first buf size is IPOIB_UD_HEAD_SIZE
118 		 */
119 		skb->tail += IPOIB_UD_HEAD_SIZE;
120 		skb->len  += length;
121 
122 		size = length - IPOIB_UD_HEAD_SIZE;
123 
124 		skb_frag_size_set(frag, size);
125 		skb->data_len += size;
126 		skb->truesize += PAGE_SIZE;
127 	} else
128 		skb_put(skb, length);
129 
130 }
131 
132 static int ipoib_ib_post_receive(struct net_device *dev, int id)
133 {
134 	struct ipoib_dev_priv *priv = netdev_priv(dev);
135 	struct ib_recv_wr *bad_wr;
136 	int ret;
137 
138 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
139 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
140 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
141 
142 
143 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
144 	if (unlikely(ret)) {
145 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
146 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
147 		dev_kfree_skb_any(priv->rx_ring[id].skb);
148 		priv->rx_ring[id].skb = NULL;
149 	}
150 
151 	return ret;
152 }
153 
154 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
155 {
156 	struct ipoib_dev_priv *priv = netdev_priv(dev);
157 	struct sk_buff *skb;
158 	int buf_size;
159 	int tailroom;
160 	u64 *mapping;
161 
162 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
163 		buf_size = IPOIB_UD_HEAD_SIZE;
164 		tailroom = 128; /* reserve some tailroom for IP/TCP headers */
165 	} else {
166 		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
167 		tailroom = 0;
168 	}
169 
170 	skb = dev_alloc_skb(buf_size + tailroom + 4);
171 	if (unlikely(!skb))
172 		return NULL;
173 
174 	/*
175 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
176 	 * header.  So we need 4 more bytes to get to 48 and align the
177 	 * IP header to a multiple of 16.
178 	 */
179 	skb_reserve(skb, 4);
180 
181 	mapping = priv->rx_ring[id].mapping;
182 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
183 				       DMA_FROM_DEVICE);
184 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
185 		goto error;
186 
187 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
188 		struct page *page = alloc_page(GFP_ATOMIC);
189 		if (!page)
190 			goto partial_error;
191 		skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
192 		mapping[1] =
193 			ib_dma_map_page(priv->ca, page,
194 					0, PAGE_SIZE, DMA_FROM_DEVICE);
195 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
196 			goto partial_error;
197 	}
198 
199 	priv->rx_ring[id].skb = skb;
200 	return skb;
201 
202 partial_error:
203 	ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
204 error:
205 	dev_kfree_skb_any(skb);
206 	return NULL;
207 }
208 
209 static int ipoib_ib_post_receives(struct net_device *dev)
210 {
211 	struct ipoib_dev_priv *priv = netdev_priv(dev);
212 	int i;
213 
214 	for (i = 0; i < ipoib_recvq_size; ++i) {
215 		if (!ipoib_alloc_rx_skb(dev, i)) {
216 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
217 			return -ENOMEM;
218 		}
219 		if (ipoib_ib_post_receive(dev, i)) {
220 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
221 			return -EIO;
222 		}
223 	}
224 
225 	return 0;
226 }
227 
228 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
229 {
230 	struct ipoib_dev_priv *priv = netdev_priv(dev);
231 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
232 	struct sk_buff *skb;
233 	u64 mapping[IPOIB_UD_RX_SG];
234 	union ib_gid *dgid;
235 
236 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
237 		       wr_id, wc->status);
238 
239 	if (unlikely(wr_id >= ipoib_recvq_size)) {
240 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
241 			   wr_id, ipoib_recvq_size);
242 		return;
243 	}
244 
245 	skb  = priv->rx_ring[wr_id].skb;
246 
247 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
248 		if (wc->status != IB_WC_WR_FLUSH_ERR)
249 			ipoib_warn(priv, "failed recv event "
250 				   "(status=%d, wrid=%d vend_err %x)\n",
251 				   wc->status, wr_id, wc->vendor_err);
252 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
253 		dev_kfree_skb_any(skb);
254 		priv->rx_ring[wr_id].skb = NULL;
255 		return;
256 	}
257 
258 	/*
259 	 * Drop packets that this interface sent, ie multicast packets
260 	 * that the HCA has replicated.
261 	 */
262 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
263 		goto repost;
264 
265 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
266 	       IPOIB_UD_RX_SG * sizeof *mapping);
267 
268 	/*
269 	 * If we can't allocate a new RX buffer, dump
270 	 * this packet and reuse the old buffer.
271 	 */
272 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
273 		++dev->stats.rx_dropped;
274 		goto repost;
275 	}
276 
277 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
278 		       wc->byte_len, wc->slid);
279 
280 	ipoib_ud_dma_unmap_rx(priv, mapping);
281 	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
282 
283 	/* First byte of dgid signals multicast when 0xff */
284 	dgid = &((struct ib_grh *)skb->data)->dgid;
285 
286 	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
287 		skb->pkt_type = PACKET_HOST;
288 	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
289 		skb->pkt_type = PACKET_BROADCAST;
290 	else
291 		skb->pkt_type = PACKET_MULTICAST;
292 
293 	skb_pull(skb, IB_GRH_BYTES);
294 
295 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
296 	skb_reset_mac_header(skb);
297 	skb_pull(skb, IPOIB_ENCAP_LEN);
298 
299 	++dev->stats.rx_packets;
300 	dev->stats.rx_bytes += skb->len;
301 
302 	skb->dev = dev;
303 	if ((dev->features & NETIF_F_RXCSUM) &&
304 			likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
305 		skb->ip_summed = CHECKSUM_UNNECESSARY;
306 
307 	napi_gro_receive(&priv->napi, skb);
308 
309 repost:
310 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
311 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
312 			   "for buf %d\n", wr_id);
313 }
314 
315 static int ipoib_dma_map_tx(struct ib_device *ca,
316 			    struct ipoib_tx_buf *tx_req)
317 {
318 	struct sk_buff *skb = tx_req->skb;
319 	u64 *mapping = tx_req->mapping;
320 	int i;
321 	int off;
322 
323 	if (skb_headlen(skb)) {
324 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
325 					       DMA_TO_DEVICE);
326 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
327 			return -EIO;
328 
329 		off = 1;
330 	} else
331 		off = 0;
332 
333 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
334 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
335 		mapping[i + off] = ib_dma_map_page(ca,
336 						 skb_frag_page(frag),
337 						 frag->page_offset, skb_frag_size(frag),
338 						 DMA_TO_DEVICE);
339 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
340 			goto partial_error;
341 	}
342 	return 0;
343 
344 partial_error:
345 	for (; i > 0; --i) {
346 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
347 
348 		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
349 	}
350 
351 	if (off)
352 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
353 
354 	return -EIO;
355 }
356 
357 static void ipoib_dma_unmap_tx(struct ib_device *ca,
358 			       struct ipoib_tx_buf *tx_req)
359 {
360 	struct sk_buff *skb = tx_req->skb;
361 	u64 *mapping = tx_req->mapping;
362 	int i;
363 	int off;
364 
365 	if (skb_headlen(skb)) {
366 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
367 		off = 1;
368 	} else
369 		off = 0;
370 
371 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
372 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
373 
374 		ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
375 				  DMA_TO_DEVICE);
376 	}
377 }
378 
379 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
380 {
381 	struct ipoib_dev_priv *priv = netdev_priv(dev);
382 	unsigned int wr_id = wc->wr_id;
383 	struct ipoib_tx_buf *tx_req;
384 
385 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
386 		       wr_id, wc->status);
387 
388 	if (unlikely(wr_id >= ipoib_sendq_size)) {
389 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
390 			   wr_id, ipoib_sendq_size);
391 		return;
392 	}
393 
394 	tx_req = &priv->tx_ring[wr_id];
395 
396 	ipoib_dma_unmap_tx(priv->ca, tx_req);
397 
398 	++dev->stats.tx_packets;
399 	dev->stats.tx_bytes += tx_req->skb->len;
400 
401 	dev_kfree_skb_any(tx_req->skb);
402 
403 	++priv->tx_tail;
404 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
405 	    netif_queue_stopped(dev) &&
406 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
407 		netif_wake_queue(dev);
408 
409 	if (wc->status != IB_WC_SUCCESS &&
410 	    wc->status != IB_WC_WR_FLUSH_ERR)
411 		ipoib_warn(priv, "failed send event "
412 			   "(status=%d, wrid=%d vend_err %x)\n",
413 			   wc->status, wr_id, wc->vendor_err);
414 }
415 
416 static int poll_tx(struct ipoib_dev_priv *priv)
417 {
418 	int n, i;
419 
420 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
421 	for (i = 0; i < n; ++i)
422 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
423 
424 	return n == MAX_SEND_CQE;
425 }
426 
427 int ipoib_poll(struct napi_struct *napi, int budget)
428 {
429 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
430 	struct net_device *dev = priv->dev;
431 	int done;
432 	int t;
433 	int n, i;
434 
435 	done  = 0;
436 
437 poll_more:
438 	while (done < budget) {
439 		int max = (budget - done);
440 
441 		t = min(IPOIB_NUM_WC, max);
442 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
443 
444 		for (i = 0; i < n; i++) {
445 			struct ib_wc *wc = priv->ibwc + i;
446 
447 			if (wc->wr_id & IPOIB_OP_RECV) {
448 				++done;
449 				if (wc->wr_id & IPOIB_OP_CM)
450 					ipoib_cm_handle_rx_wc(dev, wc);
451 				else
452 					ipoib_ib_handle_rx_wc(dev, wc);
453 			} else
454 				ipoib_cm_handle_tx_wc(priv->dev, wc);
455 		}
456 
457 		if (n != t)
458 			break;
459 	}
460 
461 	if (done < budget) {
462 		napi_complete(napi);
463 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
464 					      IB_CQ_NEXT_COMP |
465 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
466 		    napi_reschedule(napi))
467 			goto poll_more;
468 	}
469 
470 	return done;
471 }
472 
473 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
474 {
475 	struct net_device *dev = dev_ptr;
476 	struct ipoib_dev_priv *priv = netdev_priv(dev);
477 
478 	napi_schedule(&priv->napi);
479 }
480 
481 static void drain_tx_cq(struct net_device *dev)
482 {
483 	struct ipoib_dev_priv *priv = netdev_priv(dev);
484 
485 	netif_tx_lock(dev);
486 	while (poll_tx(priv))
487 		; /* nothing */
488 
489 	if (netif_queue_stopped(dev))
490 		mod_timer(&priv->poll_timer, jiffies + 1);
491 
492 	netif_tx_unlock(dev);
493 }
494 
495 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
496 {
497 	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
498 
499 	mod_timer(&priv->poll_timer, jiffies);
500 }
501 
502 static inline int post_send(struct ipoib_dev_priv *priv,
503 			    unsigned int wr_id,
504 			    struct ib_ah *address, u32 qpn,
505 			    struct ipoib_tx_buf *tx_req,
506 			    void *head, int hlen)
507 {
508 	struct ib_send_wr *bad_wr;
509 	int i, off;
510 	struct sk_buff *skb = tx_req->skb;
511 	skb_frag_t *frags = skb_shinfo(skb)->frags;
512 	int nr_frags = skb_shinfo(skb)->nr_frags;
513 	u64 *mapping = tx_req->mapping;
514 
515 	if (skb_headlen(skb)) {
516 		priv->tx_sge[0].addr         = mapping[0];
517 		priv->tx_sge[0].length       = skb_headlen(skb);
518 		off = 1;
519 	} else
520 		off = 0;
521 
522 	for (i = 0; i < nr_frags; ++i) {
523 		priv->tx_sge[i + off].addr = mapping[i + off];
524 		priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
525 	}
526 	priv->tx_wr.num_sge	     = nr_frags + off;
527 	priv->tx_wr.wr_id 	     = wr_id;
528 	priv->tx_wr.wr.ud.remote_qpn = qpn;
529 	priv->tx_wr.wr.ud.ah 	     = address;
530 
531 	if (head) {
532 		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
533 		priv->tx_wr.wr.ud.header = head;
534 		priv->tx_wr.wr.ud.hlen	 = hlen;
535 		priv->tx_wr.opcode	 = IB_WR_LSO;
536 	} else
537 		priv->tx_wr.opcode	 = IB_WR_SEND;
538 
539 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
540 }
541 
542 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
543 		struct ipoib_ah *address, u32 qpn)
544 {
545 	struct ipoib_dev_priv *priv = netdev_priv(dev);
546 	struct ipoib_tx_buf *tx_req;
547 	int hlen, rc;
548 	void *phead;
549 
550 	if (skb_is_gso(skb)) {
551 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
552 		phead = skb->data;
553 		if (unlikely(!skb_pull(skb, hlen))) {
554 			ipoib_warn(priv, "linear data too small\n");
555 			++dev->stats.tx_dropped;
556 			++dev->stats.tx_errors;
557 			dev_kfree_skb_any(skb);
558 			return;
559 		}
560 	} else {
561 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
562 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
563 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
564 			++dev->stats.tx_dropped;
565 			++dev->stats.tx_errors;
566 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
567 			return;
568 		}
569 		phead = NULL;
570 		hlen  = 0;
571 	}
572 
573 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
574 		       skb->len, address, qpn);
575 
576 	/*
577 	 * We put the skb into the tx_ring _before_ we call post_send()
578 	 * because it's entirely possible that the completion handler will
579 	 * run before we execute anything after the post_send().  That
580 	 * means we have to make sure everything is properly recorded and
581 	 * our state is consistent before we call post_send().
582 	 */
583 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
584 	tx_req->skb = skb;
585 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
586 		++dev->stats.tx_errors;
587 		dev_kfree_skb_any(skb);
588 		return;
589 	}
590 
591 	if (skb->ip_summed == CHECKSUM_PARTIAL)
592 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
593 	else
594 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
595 
596 	if (++priv->tx_outstanding == ipoib_sendq_size) {
597 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
598 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
599 			ipoib_warn(priv, "request notify on send CQ failed\n");
600 		netif_stop_queue(dev);
601 	}
602 
603 	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
604 		       address->ah, qpn, tx_req, phead, hlen);
605 	if (unlikely(rc)) {
606 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
607 		++dev->stats.tx_errors;
608 		--priv->tx_outstanding;
609 		ipoib_dma_unmap_tx(priv->ca, tx_req);
610 		dev_kfree_skb_any(skb);
611 		if (netif_queue_stopped(dev))
612 			netif_wake_queue(dev);
613 	} else {
614 		dev->trans_start = jiffies;
615 
616 		address->last_send = priv->tx_head;
617 		++priv->tx_head;
618 		skb_orphan(skb);
619 
620 	}
621 
622 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
623 		while (poll_tx(priv))
624 			; /* nothing */
625 }
626 
627 static void __ipoib_reap_ah(struct net_device *dev)
628 {
629 	struct ipoib_dev_priv *priv = netdev_priv(dev);
630 	struct ipoib_ah *ah, *tah;
631 	LIST_HEAD(remove_list);
632 	unsigned long flags;
633 
634 	netif_tx_lock_bh(dev);
635 	spin_lock_irqsave(&priv->lock, flags);
636 
637 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
638 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
639 			list_del(&ah->list);
640 			ib_destroy_ah(ah->ah);
641 			kfree(ah);
642 		}
643 
644 	spin_unlock_irqrestore(&priv->lock, flags);
645 	netif_tx_unlock_bh(dev);
646 }
647 
648 void ipoib_reap_ah(struct work_struct *work)
649 {
650 	struct ipoib_dev_priv *priv =
651 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
652 	struct net_device *dev = priv->dev;
653 
654 	__ipoib_reap_ah(dev);
655 
656 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
657 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
658 				   round_jiffies_relative(HZ));
659 }
660 
661 static void ipoib_ib_tx_timer_func(unsigned long ctx)
662 {
663 	drain_tx_cq((struct net_device *)ctx);
664 }
665 
666 int ipoib_ib_dev_open(struct net_device *dev)
667 {
668 	struct ipoib_dev_priv *priv = netdev_priv(dev);
669 	int ret;
670 
671 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
672 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
673 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
674 		return -1;
675 	}
676 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
677 
678 	ret = ipoib_init_qp(dev);
679 	if (ret) {
680 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
681 		return -1;
682 	}
683 
684 	ret = ipoib_ib_post_receives(dev);
685 	if (ret) {
686 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
687 		ipoib_ib_dev_stop(dev, 1);
688 		return -1;
689 	}
690 
691 	ret = ipoib_cm_dev_open(dev);
692 	if (ret) {
693 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
694 		ipoib_ib_dev_stop(dev, 1);
695 		return -1;
696 	}
697 
698 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
699 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
700 			   round_jiffies_relative(HZ));
701 
702 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
703 		napi_enable(&priv->napi);
704 
705 	return 0;
706 }
707 
708 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
709 {
710 	struct ipoib_dev_priv *priv = netdev_priv(dev);
711 	u16 pkey_index = 0;
712 
713 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
714 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
715 	else
716 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
717 }
718 
719 int ipoib_ib_dev_up(struct net_device *dev)
720 {
721 	struct ipoib_dev_priv *priv = netdev_priv(dev);
722 
723 	ipoib_pkey_dev_check_presence(dev);
724 
725 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
726 		ipoib_dbg(priv, "PKEY is not assigned.\n");
727 		return 0;
728 	}
729 
730 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
731 
732 	return ipoib_mcast_start_thread(dev);
733 }
734 
735 int ipoib_ib_dev_down(struct net_device *dev, int flush)
736 {
737 	struct ipoib_dev_priv *priv = netdev_priv(dev);
738 
739 	ipoib_dbg(priv, "downing ib_dev\n");
740 
741 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
742 	netif_carrier_off(dev);
743 
744 	/* Shutdown the P_Key thread if still active */
745 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
746 		mutex_lock(&pkey_mutex);
747 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
748 		cancel_delayed_work(&priv->pkey_poll_task);
749 		mutex_unlock(&pkey_mutex);
750 		if (flush)
751 			flush_workqueue(ipoib_workqueue);
752 	}
753 
754 	ipoib_mcast_stop_thread(dev, flush);
755 	ipoib_mcast_dev_flush(dev);
756 
757 	ipoib_flush_paths(dev);
758 
759 	return 0;
760 }
761 
762 static int recvs_pending(struct net_device *dev)
763 {
764 	struct ipoib_dev_priv *priv = netdev_priv(dev);
765 	int pending = 0;
766 	int i;
767 
768 	for (i = 0; i < ipoib_recvq_size; ++i)
769 		if (priv->rx_ring[i].skb)
770 			++pending;
771 
772 	return pending;
773 }
774 
775 void ipoib_drain_cq(struct net_device *dev)
776 {
777 	struct ipoib_dev_priv *priv = netdev_priv(dev);
778 	int i, n;
779 
780 	/*
781 	 * We call completion handling routines that expect to be
782 	 * called from the BH-disabled NAPI poll context, so disable
783 	 * BHs here too.
784 	 */
785 	local_bh_disable();
786 
787 	do {
788 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
789 		for (i = 0; i < n; ++i) {
790 			/*
791 			 * Convert any successful completions to flush
792 			 * errors to avoid passing packets up the
793 			 * stack after bringing the device down.
794 			 */
795 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
796 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
797 
798 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
799 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
800 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
801 				else
802 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
803 			} else
804 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
805 		}
806 	} while (n == IPOIB_NUM_WC);
807 
808 	while (poll_tx(priv))
809 		; /* nothing */
810 
811 	local_bh_enable();
812 }
813 
814 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
815 {
816 	struct ipoib_dev_priv *priv = netdev_priv(dev);
817 	struct ib_qp_attr qp_attr;
818 	unsigned long begin;
819 	struct ipoib_tx_buf *tx_req;
820 	int i;
821 
822 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
823 		napi_disable(&priv->napi);
824 
825 	ipoib_cm_dev_stop(dev);
826 
827 	/*
828 	 * Move our QP to the error state and then reinitialize in
829 	 * when all work requests have completed or have been flushed.
830 	 */
831 	qp_attr.qp_state = IB_QPS_ERR;
832 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
833 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
834 
835 	/* Wait for all sends and receives to complete */
836 	begin = jiffies;
837 
838 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
839 		if (time_after(jiffies, begin + 5 * HZ)) {
840 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
841 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
842 
843 			/*
844 			 * assume the HW is wedged and just free up
845 			 * all our pending work requests.
846 			 */
847 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
848 				tx_req = &priv->tx_ring[priv->tx_tail &
849 							(ipoib_sendq_size - 1)];
850 				ipoib_dma_unmap_tx(priv->ca, tx_req);
851 				dev_kfree_skb_any(tx_req->skb);
852 				++priv->tx_tail;
853 				--priv->tx_outstanding;
854 			}
855 
856 			for (i = 0; i < ipoib_recvq_size; ++i) {
857 				struct ipoib_rx_buf *rx_req;
858 
859 				rx_req = &priv->rx_ring[i];
860 				if (!rx_req->skb)
861 					continue;
862 				ipoib_ud_dma_unmap_rx(priv,
863 						      priv->rx_ring[i].mapping);
864 				dev_kfree_skb_any(rx_req->skb);
865 				rx_req->skb = NULL;
866 			}
867 
868 			goto timeout;
869 		}
870 
871 		ipoib_drain_cq(dev);
872 
873 		msleep(1);
874 	}
875 
876 	ipoib_dbg(priv, "All sends and receives done.\n");
877 
878 timeout:
879 	del_timer_sync(&priv->poll_timer);
880 	qp_attr.qp_state = IB_QPS_RESET;
881 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
882 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
883 
884 	/* Wait for all AHs to be reaped */
885 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
886 	cancel_delayed_work(&priv->ah_reap_task);
887 	if (flush)
888 		flush_workqueue(ipoib_workqueue);
889 
890 	begin = jiffies;
891 
892 	while (!list_empty(&priv->dead_ahs)) {
893 		__ipoib_reap_ah(dev);
894 
895 		if (time_after(jiffies, begin + HZ)) {
896 			ipoib_warn(priv, "timing out; will leak address handles\n");
897 			break;
898 		}
899 
900 		msleep(1);
901 	}
902 
903 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
904 
905 	return 0;
906 }
907 
908 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
909 {
910 	struct ipoib_dev_priv *priv = netdev_priv(dev);
911 
912 	priv->ca = ca;
913 	priv->port = port;
914 	priv->qp = NULL;
915 
916 	if (ipoib_transport_dev_init(dev, ca)) {
917 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
918 		return -ENODEV;
919 	}
920 
921 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
922 		    (unsigned long) dev);
923 
924 	if (dev->flags & IFF_UP) {
925 		if (ipoib_ib_dev_open(dev)) {
926 			ipoib_transport_dev_cleanup(dev);
927 			return -ENODEV;
928 		}
929 	}
930 
931 	return 0;
932 }
933 
934 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
935 				enum ipoib_flush_level level)
936 {
937 	struct ipoib_dev_priv *cpriv;
938 	struct net_device *dev = priv->dev;
939 	u16 new_index;
940 
941 	mutex_lock(&priv->vlan_mutex);
942 
943 	/*
944 	 * Flush any child interfaces too -- they might be up even if
945 	 * the parent is down.
946 	 */
947 	list_for_each_entry(cpriv, &priv->child_intfs, list)
948 		__ipoib_ib_dev_flush(cpriv, level);
949 
950 	mutex_unlock(&priv->vlan_mutex);
951 
952 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
953 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
954 		return;
955 	}
956 
957 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
958 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
959 		return;
960 	}
961 
962 	if (level == IPOIB_FLUSH_HEAVY) {
963 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
964 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
965 			ipoib_ib_dev_down(dev, 0);
966 			ipoib_ib_dev_stop(dev, 0);
967 			if (ipoib_pkey_dev_delay_open(dev))
968 				return;
969 		}
970 
971 		/* restart QP only if P_Key index is changed */
972 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
973 		    new_index == priv->pkey_index) {
974 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
975 			return;
976 		}
977 		priv->pkey_index = new_index;
978 	}
979 
980 	if (level == IPOIB_FLUSH_LIGHT) {
981 		ipoib_mark_paths_invalid(dev);
982 		ipoib_mcast_dev_flush(dev);
983 	}
984 
985 	if (level >= IPOIB_FLUSH_NORMAL)
986 		ipoib_ib_dev_down(dev, 0);
987 
988 	if (level == IPOIB_FLUSH_HEAVY) {
989 		ipoib_ib_dev_stop(dev, 0);
990 		ipoib_ib_dev_open(dev);
991 	}
992 
993 	/*
994 	 * The device could have been brought down between the start and when
995 	 * we get here, don't bring it back up if it's not configured up
996 	 */
997 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
998 		if (level >= IPOIB_FLUSH_NORMAL)
999 			ipoib_ib_dev_up(dev);
1000 		ipoib_mcast_restart_task(&priv->restart_task);
1001 	}
1002 }
1003 
1004 void ipoib_ib_dev_flush_light(struct work_struct *work)
1005 {
1006 	struct ipoib_dev_priv *priv =
1007 		container_of(work, struct ipoib_dev_priv, flush_light);
1008 
1009 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1010 }
1011 
1012 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1013 {
1014 	struct ipoib_dev_priv *priv =
1015 		container_of(work, struct ipoib_dev_priv, flush_normal);
1016 
1017 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1018 }
1019 
1020 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1021 {
1022 	struct ipoib_dev_priv *priv =
1023 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1024 
1025 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1026 }
1027 
1028 void ipoib_ib_dev_cleanup(struct net_device *dev)
1029 {
1030 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1031 
1032 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1033 
1034 	ipoib_mcast_stop_thread(dev, 1);
1035 	ipoib_mcast_dev_flush(dev);
1036 
1037 	ipoib_transport_dev_cleanup(dev);
1038 }
1039 
1040 /*
1041  * Delayed P_Key Assigment Interim Support
1042  *
1043  * The following is initial implementation of delayed P_Key assigment
1044  * mechanism. It is using the same approach implemented for the multicast
1045  * group join. The single goal of this implementation is to quickly address
1046  * Bug #2507. This implementation will probably be removed when the P_Key
1047  * change async notification is available.
1048  */
1049 
1050 void ipoib_pkey_poll(struct work_struct *work)
1051 {
1052 	struct ipoib_dev_priv *priv =
1053 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1054 	struct net_device *dev = priv->dev;
1055 
1056 	ipoib_pkey_dev_check_presence(dev);
1057 
1058 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1059 		ipoib_open(dev);
1060 	else {
1061 		mutex_lock(&pkey_mutex);
1062 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1063 			queue_delayed_work(ipoib_workqueue,
1064 					   &priv->pkey_poll_task,
1065 					   HZ);
1066 		mutex_unlock(&pkey_mutex);
1067 	}
1068 }
1069 
1070 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1071 {
1072 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1073 
1074 	/* Look for the interface pkey value in the IB Port P_Key table and */
1075 	/* set the interface pkey assigment flag                            */
1076 	ipoib_pkey_dev_check_presence(dev);
1077 
1078 	/* P_Key value not assigned yet - start polling */
1079 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1080 		mutex_lock(&pkey_mutex);
1081 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1082 		queue_delayed_work(ipoib_workqueue,
1083 				   &priv->pkey_poll_task,
1084 				   HZ);
1085 		mutex_unlock(&pkey_mutex);
1086 		return 1;
1087 	}
1088 
1089 	return 0;
1090 }
1091